The measurement of raw water turbidity is essential to determine its degree of potabilization. However, an erroneous measurement of this treatment can cause the water to be of insufficient quality to be purified, generating health problems for consumers. Unfortunately, in rural treatment plants in Colombia, it is common to find this type of deficiency because they use rudimentary techniques to evaluate turbidity, making it impossible to guarantee sufficient quality for water to be considered fit for human consumption.
This article proposes the use of machine learning models in the treatment plants from the municipalities of Santander de Quilichao, Timbío, and Mercaderes in the department of Cauca (Colombia), in order to estimate the turbidity value of raw water considering only available data such as pH, temperature, vapor pressure, and precipitation captured manually by plant's operators.
To develop these modeles, Linear Regression, Random Forest Regressor, Kneighbors Regressor, and Extra Trees Regressor algorithms were trained with data provided by the Aquarisc project and the Colombian Institute of Hydrology, Meteorology and Environment Studies (IDEAM). The study determined that the best-performing algorithm in this context was the Random Forest Regressor. This algorithm had the best RMSE among all the models and was also the one that best fitted the situation of the treatment plants studied. Furthermore, this model did not consider outliers and obtained a result of an RMSE of 20.98 and 3.49 for the training and test data sets, respectively. Finally, it was determined that this algorithm could acceptably estimate the water's turbidity and may support the operators in the decision-making process for applying an adequate treatment to the drinking water.